Integration guide
Home Assistant Integration
Open-source smart home platform that can integrate AI assistants, voice control, automations, and local model workflows.
Recommended path
Home Assistant local AI quick start
- 1. Run Ollama with a tool-capable chat modelStart Ollama, pull a model that explicitly supports tools, and confirm that the model answers a basic prompt before connecting Home Assistant.
- 2. Add the official Ollama integrationIn Home Assistant, add Ollama, enter the server URL such as http://localhost:11434, and select the model you already pulled.
- 3. Expose a small, deliberate entity setEntity control is experimental. Begin with fewer than 25 entities, test read-only questions first, and enable control only after the model handles names and areas reliably.
Best for
Builders who want AI-assisted smart home control, local voice experiments, and automation workflows.
Model support
The official Ollama integration adds a conversation agent backed by your own Ollama server: you supply the server URL and pick a model from the Ollama library. Controlling the house, rather than just chatting, additionally requires a model that supports tool calling.
Choose models by role
Conversation and entity control
Tool-capable Ollama chat model
Required for controlling Home Assistant. A chat model without tool support can converse but cannot call the Assist API to act on entities.
Optional speech-to-text
Whisper Large V3
Transcribes voice input in a local voice pipeline. It does not replace the conversation model and does not control entities.
Open the model profile →How to use this integration
- Replace the built-in voice assistant with a local conversation agent so recordings never leave the house
- Ask about house state in natural language instead of memorising entity names
- Control a deliberately small set of exposed entities by voice, using a tool-capable model
- Prototype automation ideas against a local model before committing them to YAML
Connecting a local model
Add the Ollama integration and point it at your server, for example http://localhost:11434, then choose a pulled model. The default context window is 8k tokens. Home Assistant documents this as an experimental capability that hands the model access to the Assist API.
Tradeoffs
Only models that support tools can actually control Home Assistant; everything else is limited to conversation. The docs recommend exposing fewer than 25 entities to a local model and warn that smaller models make more mistakes, so a Pi-class setup will misfire on a large house. The integration also does not work with sentence triggers.